Outlier Robust Extreme Learning Machine for Multi-Target Regression

2019·arXiv

Abstract

Abstract

The popularity of algorithms based on Extreme Learning Machine (ELM), which can be used to train Single Layer Feedforward Neural Networks (SLFN), has increased in the past years. They have been successfully applied to a wide range of classification and regression tasks. The most commonly used methods are the ones based on minimizing the norm of the error, which is not suitable to deal with outliers, essentially in regression tasks. The use of norm was proposed in Outlier Robust ELM (OR-ELM), which is defined to one-dimensional outputs. In this paper, we generalize OR-ELM to deal with multi-target regression problems, using the error norm and the Elastic Net theory, which can result in a more sparse network, resulting in our method, Generalized Outlier Robust ELM (GOR-ELM). We use Alternating Direction Method of Multipliers (ADMM) to solve the resulting optimization problem. An incremental version of GOR-ELM is also proposed. We chose 15 public real-world multi-target regression datasets to test our methods. Our conducted experiments show that they are statistically better than other ELM-based techniques, when considering data contaminated with outliers, and equivalent to them, otherwise.

norm, Extreme Learning Machine, Regularization,

Multi-target Regression, Robust to Outliers, Alternating direction method of

multipliers

1. Introduction

In the past years, a fast algorithm called Extreme Learning Machine (ELM) was proposed by Huang et al. (2004, 2006b). It is used to train Single Layer Feedforward Networks (SLFN), as shown in Fig. 1, where part of the network parameters (and ) are randomly generated, and the remaining () are found using labeled data and a closed-form solution.

Figure 1: Illustration of an SLFN architecture with hidden nodes and m outputs (Inaba et al., 2018).

Due to its simplicity and speed, ELM gained popularity and has been applied in a wide range of applications such as computer vision and time series analysis (Huang et al., 2015), achieving good performances, generally better than classifiers such as Support Vector Machines (SVMs) (Huang et al., 2015).

Variants of ELM were also proposed, turning possible to deal with sequential data (where new samples arrive over time) using Online Sequential ELM (OSELM) (Liang et al., 2006), or increasing the SLFN hidden node number, using Incremental ELM (I-ELM) (Huang et al., 2006a).

In ELM and OS-ELM, the number of nodes in the hidden layer needs to be well-chosen to avoid overfitting and underfitting (Deng et al., 2009). To overcome this problem, an ELM-based algorithm using ridge regression was proposed by Deng et al. (2009). Although it achieves good results, the resulting network is dense and might suffer from memory and computing limitations.

Mart´ınez-Mart´ınez et al. (2011) extended Deng et al. (2009) work, proposing an algorithm named Regularized ELM (R-ELM), which can select an architecture based on the problem. By using the Elastic Net theory, R-ELM can prune some hidden nodes when dealing with a one-dimensional output.

The aforementioned methods were defined considering only one output node, with optimization problems minimizing vector norms, but they can be adapted to multi-dimensional tasks by considering each output separately. To deal with these tasks, Generalized Regularized ELM (GR-ELM) was proposed by Inaba et al. (2018), which generalizes R-ELM, using matrix norms in its objective function, replacing and vector norms by Frobenius and matrix norms, respectively.

One common characteristic of these methods is the use of or Frobenius norm of the prediction error in each objective function. As pointed by Zhang & Luo (2015), these approaches have some drawbacks when the application suffers from outliers, which is common in real-world tasks, since the model can be biased towards them.

To deal with outliers, Outlier Robust ELM (OR-ELM) was proposed by Zhang & Luo (2015), which considers the norm of the prediction error in its objective function, achieving better results in the presence of outliers in regression tasks, when compared with other algorithms. However, OR-ELM was also defined only to one-dimensional outputs.

In this paper, we generalize the OR-ELM to multi-target regression problems. We use the same model considered in GR-ELM, replacing Frobenius norm by norm of the prediction error, which can be interpreted as an extension of the vector norm. When considering outputs with only one dimension and using only the ridge penalty, our method is the same as OR-ELM. We use a three-block Alternating Direction Method of Multipliers (ADMM) (Chen et al., 2016) to solve our optimization problem, which is a simple but powerful algorithm. We also propose an incremental version of this method, which can increase the number of nodes efficiently if the desired error was not obtained.

Our methods were tested in 15 public real-world datasets, which were contaminated with outliers to verify the robustness of them, using the average relative root mean square error (aRRMSE) as metric. We compared our methods with similar algorithms based on ELM.

The paper is organized as follows. In Section 2, we review some ELM-based algorithms. We describe the proposed methods in Section 3. Experiments and results are presented in Section 4, and Section 5 concludes the paper.

2. Related Work

In this section, we review Extreme Learning Machine and some of its variants, in specific regularized and incremental ones.

2.1. ELM

An extremely fast algorithm to train an SLFN with hidden nodes was proposed in Huang et al. (2004). This algorithm was called ELM and considers a dataset with N distinct and labeled samples (), where = [and . The SLFN estimation of is modeled

as

where = [is the weight vector connecting the i-th hidden node and the input nodes. = [is the weight vector connecting the hidden nodes and the output node, is the bias of the i-th hidden node, and ) is the activation function of the SLFN.

Assuming that an SLFN with nodes can approximate all N samples with zero error, i.e, ˆ= 1, . . . , N, we can rewrite Eq. 1 as:

where t = [and

where A = [, . . . , = [ν, . . . , ν and X = [, . . . , .

The ELM solution, ˆ, is the smallest norm least-square solution of the linear

system given by Eq. 2, i.e.,

which has a closed-form solution:

where is the Moore-Penrose generalized inverse (Rao & Mitra, 1971) of H.

2.2. Regularized ELM

Although ELM has shown good results in several applications, the right choice of must be made in order to obtain a good performance, avoiding overfitting and underfitting. Bartlett (1998) showed that models whose parameters have smaller norms are capable of achieving better generalization. On this note, Deng et al. (2009) introduced R-ELM for SLFN with sigmoid additive nodes, where an norm of is added in the optimization problem of Eq. 4. Huang et al. (2012) extends this work to different types of hidden nodes and activation functions, as well as kernels.

Other types of regularization in the ELM optimization problem were considered by Mart´ınez-Mart´ınez et al. (2011). Then, R-ELM can be described in

a generalized way by

where C and are regularization parameters.

The optimization problem of Eq. 6 uses the Elastic Net penalty, which is a trade-off between the ridge regression (= 0), where only the norm is considered, and the lasso penalty (= 1), where only the norm is considered. Since it uses the norm, the Elastic Net method is also capable of reducing the network size, pruning output weights while maintaining generalization.

When considering only the penalty (= 0) with = 0, as done by Deng

et al. (2009) and Huang et al. (2012), the solution of R-ELM is given by

2.3. Outlier Robust ELM

Recently, the norm has been applied to improve the performance of methods in the presence of outliers (Zhang & Luo, 2015). This can be achieved in

ELM by using it in the loss function, i.e.,

instead of using the usual norm.

Supported on this observation, Zhang & Luo (2015) proposed the OR-ELM algorithm, which finds the solution of the following optimization problem:

This optimization problem can be solved by using the Augmented Lagrange Multiplier (ALM) method, as suggested by Zhang & Luo (2015).

2.4. Generalized Regularized ELM

A limitation of ELM, R-ELM and OR-ELM is that they are defined only for one-dimensional outputs. Although we can consider each output separately in multi-dimensional tasks, their solutions are not capable of capturing relations between the multiple outputs. This means that we can only prune a hidden node if and only if its weights are pruned in all outputs simultaneously, which can be difficult.

To capture those relations, matrix norms can be used in the ELM optimization problem. Considering this, the GR-ELM method was proposed by Inaba et al. (2018). This method considers a optimization problem similar to Eq. 6, using matrices and replacing and norms by Frobenius and norms, respectively.

Considering a dataset with multi-dimensional outputs, with N distinct sam-

ples (), where = [x, x, . . . , xand = [t, t, . . . , t,

we can extend Eq. 2:

where T = [and B is the weight matrix connecting the SLFN hidden and output layers, i.e., B = [. Then, the opti-

mization problem of GR-ELM is given by

where and are regularization parameters, is the Frobenius norm of a A = [matrix with n rows and m columns, where is

the i-th row of A, defined as

where ) is the trace of and is the norm of A, defined

as

This optimization problem can be solved by using the Alternating Direction Method of Multipliers (ADMM) (Boyd, 2010) method, as suggested by Inaba et al. (2018). Algorithm 1 summarizes the GR-ELM method, where is an

ADMM parameter.

2.5. Incremental ELM

Due to its simplicity and low computational cost, the ELM algorithm was very popular after it was proposed. However, it has two main challenges (Feng et al., 2009): reducing its complexity to deal with large datasets and models with many nodes; and choosing the optimal hidden node number (usually a trial and error method is used).

To deal with those problems, Huang et al. (2006b) proposed an algorithm named I-ELM to train an SLFN adding one node at a time, reducing the memory cost (since in each iteration, it only works with one node) of training an SLFN. Considering a dataset with one-dimensional targets, the algorithm starts with

N = 0 and considers that the residual error is equal to the targets: = t.

When the j-th node is added, its input weights and bias are randomly generated, and then its output of every training sample is calculated as = [. Finally, the output weight of the new

node is calculated using

and the residual error is updated using

Following this algorithm, it is possible to achieve good approximations of functions, with good generalization performance (Huang et al., 2006b). Furthermore, by using an expected learning accuracy as a stopping criterion, it is possible to find a good number of hidden nodes without using the trial and error approach.

One disadvantage of the I-ELM algorithm is that only one hidden node can be added at a time. An incremental method that was capable of adding a group of hidden nodes at the same time in an SLFN was proposed by Feng et al. (2009). This method is called Error Minimization ELM (EM-ELM) and takes advantage of the Schur complement (Petersen & Pedersen, 2007) to update the generalized inverse of the matrix , which is updated in each k-th iteration.

Assuming we already have an SLFN, with one output and hidden nodes, trained using the ELM algorithm, we used a matrix to calculate the output weights . When new nodes are added to the network, its new hidden output matrix can be written as:

where

According to Feng et al. (2009), it is possible to update the output weights

where and are given by

where I is an identity matrix and

respectively.

Note that if is stored, we can save some computations when adding the new nodes in the k + 1 iteration, reducing the algorithm computational time.

2.6. Incremental Regularized ELM

The I-ELM and EM-ELM algorithms are capable of increasing the hidden node number of an SLFN trained using the original ELM algorithm. However, these methods have a problem in some applications where the initial hidden layer output matrix is rank deficient, and the accuracy of its computations may be compromised (Xu et al., 2016). Also, they inherit ELM problems and sometimes cannot achieve the expected testing accuracy.

The overfitting problem of ELM can be solved by using the structural risk minimization principle, which was used in the R-ELM method (Deng et al., 2009; Mart´ınez-Mart´ınez et al., 2011; Huang et al., 2012). However, the hidden node number of an SLFN trained with the original R-ELM method is a hyperparameter and cannot be increased efficiently.

To solve this problem, the Incremental Regularized ELM (IR-ELM) was proposed (Xu et al., 2016), which is an algorithm capable of increasing the hidden node number of an SLFN trained using the R-ELM algorithm (where the considered structural risk is the norm, with = 0 and = 0). IR-ELM considers that by adding one node, its new hidden output matrix is given

by = [], where = [+ and the new output weights of the SLFN can be calculated as

We can rewrite as

According to Xu et al. (2016), can be updated using:

where and are given by

and

respectively.

By using Eq. 22, we can find the new network output weight () effi-ciently, using known information, which is usually faster than training a new and larger SLFN. According to Zhang & Luo (2015), methods that considers the error norm can suffer in the presence of outliers, which is the case of IR-ELM.

3. Proposed Method

In this section, we present our proposed methods. We first generalize ORELM to multi-target regression problems, presenting its ADMM update rules. This method is named Generalized Outlier Robust ELM (GOR-ELM). We also present an incremental version of this generalization.

3.1. GOR-ELM

Multi-target regression (MTR) refers to the prediction problems of multiple real-valued outputs. Such problems occur in different fields, including ecological modeling, economy, energy, data mining, and medical image (Ghosn & Bengio, 1997; Dˇzeroski et al., 2000; Kocev et al., 2009; Wang et al., 2015; Zhen et al., 2016; Spyromitros-Xioufis et al., 2016).

A natural way to treat MTR problems is to obtain a model for each output, separately. However, one of the main MTR challenges is contemplating the output relationship, besides the non-linear input-output mapping, on modeling (Zhang & Yeung, 2012; Zhen et al., 2017).

In the age of big data, MTR problems are becoming more and more common in the massive amount of accumulated data. However, inherent to this immense volume of data, other problems arise. Outliers are becoming more frequent due to different causes, such as instrumentation or human error (Zhang & Luo, 2015; Hodge & Austin, 2004). Although modeling the complex relationship between outputs and input-output on MTR problems are well studied, little has been done regarding outliers on MTR problems.

The use of norm to achieve robustness to outliers in regression problems is a common practice in the literature (Zhang & Luo, 2015; Xu et al., 2013; Chen et al., 2017). The idea of GOR-ELM is to extend the OR-ELM to MTR problems. Nevertheless, instead of treating each output separately, possible output relationships are considered on GOR-ELM through its regularization. Therefore, GOR-ELM is especially suitable for problems where output outliers may occur in a structuredway.

Thus, our approach is an extension of OR-ELM where matrix norms are considered, instead of vector ones. We replace and norms by and Frobenius norms, respectively. We can also view our approach as an extension of GR-ELM (Eq. 11), replacing the Frobenius norm of the error by its norm. According to Ma et al. (2017) and Mandanas & Kotropoulos (2018), the also offers robustness to outliers, as well as norm, which implies that GOR-ELM returns a robust network.

The following optimization problem is proposed

The optimization problem of GOR-ELM (Eq. 25) is equivalent to the fol-

lowing constrained problem

where the objective function is separable on E, B, and Z. Similar to GRELM, we use ADMM to solve the optimization problem. On each iteration of ADMM, the algorithm performs an alternated minimization of the augmented Lagrangian over E, B, and Z.

Note that we can rewrite Eq. 26 as

where ˜, , and = . According to Chen et al. (2016), the optimization problem

established in Eq. 27 can be solved using the 3-block ADMM algorithm, which

does not necessarily converge. A sufficient condition to its convergence is that at least one of the following statements is true: ˜, ˜or ˜. In our method, we can check that ˜is true. Then, GOR-ELM algorithm converges. For more information, see Chen et al. (2016).

The augmented Lagrangian of the Eq. 27 is

where Y is the Lagrange multiplier, and 0 is a regularization parameter.

Using the scaled dual variable U = (1, we can rewrite the augmented

Lagrangian as

Since the Frobenius norm is separable, and by the equivalence of Eq. 26 and

Eq. 27, the augmented Lagrangian (Eq. 29) of GOR-ELM can be written as

where , with and .

At iteration k, ADMM consists of the following update rules for

1. , we have the following subproblem

2. , we have the following subproblem

3. , we have the following subproblem

4. , the Lagrange multiplier is updated by

Algorithm 2 resumes the GOR-ELM method. Fig. 2 shows the proposed

algorithm flowchart.

Initialization: and k = 0.

Generate random weights A and biases .

Calculate the output of nodes H using Eq. 3.

repeat

:= arg min+ +

:= arg minHBE + U:= U+

k = k+1

stopping criterion not meet;

Figure 2: Proposed algorithm flowchart.

3.2. Stopping Criterion

We follow the suggestions given in Chen et al. (2016) and Boyd & Vanden- berghe (2009) for stopping criterion. We terminate the algorithm when primal

and dual residuals, given by

and

respectively, satisfy

and

The tolerances 0 and 0 are set using the absolute and relative

criterion

where 0 and 0 are the absolute and relative tolerance, respectively.

3.3. Algorithm Complexity

When analyzing Algorithm 2, we can see that the most expensive steps are the B, Z and E updates, using Eqs. 33, 35 and 38, respectively. When considering Eq. 33, the most expensive operation is the matrix inversion, which has complexity O( + ), according to Akusok et al. (2015). We can consider this as the complexity of each iteration, since Eqs. 35 and 38 are composed of the block soft-thresholding operation, which is a less complex operation.

3.4. IGOR-ELM

In some applications, we can choose an initial SLFN that achieves an undesired error when trained using any ELM-based algorithm. This can mean that there is still room for improvement in the model generalization. As discussed in Section 2.5, one of the approaches that can improve the performance of an SLFN is to increase its number of hidden nodes.

To train this larger model, we can simply use the same algorithm using a new, untrained, SLFN with more nodes. However, this procedure is not efficient if we already have a trained model, since we would discard all learned information. Thus, if the initial model efficiency is not sufficient, an approach to increase the number of hidden nodes taking advantage of the previous knowledge will be desirable.

In the literature, there are some ELM-based algorithms, such as I-ELM (Huang et al., 2006b), EM-ELM (Feng et al., 2009) and IR-ELM (Xu et al., 2016), that are capable of adding more nodes to an existing SLFN that was trained using an ELM algorithm, taking advantage of known information to find the new model in less time. However, these methods usually suffer in the presence of outliers.

In this paper, we also propose the Incremental Generalized Outlier Robust ELM (IGOR-ELM). As the name suggests, it is an incremental version of GORELM, where it is possible to add a new group of hidden nodes at each iteration to an existing SLFN and is robust to outliers. This algorithm solves almost the same optimization problem of GOR-ELM (Eq. 25), taking advantage of a known model to train a larger one.

Since it uses the elastic net theory, GOR-ELM can prune hidden nodes. However, its optimization problem considers a trade-off between pruning nodes and the impact of this action in the model error. This implies that if no nodes were pruned, there is still room for improvement, and we can use IGOR-ELM to increase the model performance.

IGOR-ELM increases the network size until some stopping criterion is achieved. Every time a larger model is trained, previous knowledge is used as a starting point to the ADMM algorithm, used to solve the GOR-ELM method (Algorithm 2).

The starting point of each IGOR-ELM iteration is composed of increasing the dimensions of matrices with a direct relationship to the number of nodes (e.g., , and Z), using new values (e.g., zero-valued matrices) in these new elements. Then, random weights and biases of the new nodes are generated. The dimension of H also increases, which implies that a larger matrix inversion must be done, which may increase the training time of each iteration.

When adding the s-th batch of nodes in an SLFN trained with IGOR-ELM, we can update the inverse of+ using its Schur complement (Pe- tersen & Pedersen, 2007) efficiently, if =+ is stored:

where

and = [].

Then, by adjusting variables whose dimension depends on the number of nodes and updating+ , the same algorithm of GOR-ELM is used and its solution is found by using its update rules.

As the stopping criterion of the IGOR-ELM algorithm, we can chose hyperparameters such as the maximum number of total hidden nodes, an expected value for the model efficiency (e.g. a metric), or the ratio of pruned nodes. Algorithm 3 resumes the IGOR-ELM method.

4. Experiments

To evaluate the proposed GOR-ELM and IGOR-ELM methods, we selected 15 public real-word datasets (Table 1) for multi-target regression (MTR) from the Mulan Java library (Tsoumakas et al., 2010), excluding the ones with missing data. The datasets are randomly divided into two parts: a training subset with 2/3 of the samples and a testing one with the remaining.

Table 1: Information about multi-target regression datasets.

We consider that a sample is an outlier based on boxplot analysis, as defined by Tukey (1977). Considering and as the first and third quartile, and as the interquartile range of the boxplot of an attribute, a sample is considered an outlier if it is located in one of the following two intervals:

[QIQR, Q1.5 IQR] and [Q+ 1.5 IQR, Q+ 3 IQR].

To evaluate the outlier robustness of our methods, we contaminated the MTR datasets using the following procedure: a subset of samples of the training set is randomly selected, which size is defined according to the outlier ratio; then each attribute of the corresponding targets of such subset are replaced by random values from ) or + 1+ 3 ). We tested our algorithms using outlier ratios of 20% and 40%, along with the uncontaminated datasets. This procedure was only used in the training subset.

The inputs and targets were normalized to have values in the interval [1] using each respective minimum and maximum values on the training dataset. This normalization was also used in the testing dataset, using the same factors. Since R-ELM, OR-ELM and IR-ELM methods were defined to one-dimensional outputs and we used multi-target datasets, they were adapted to deal with each dimension separately, where each is found considering only the i-th output,

i.e,

and the output matrix ˆB is constructed:

Our tests were conducted in an Ubuntu 18.04 Linux computer with Intel Corei7-8700K CPU with 32 GB of RAM, using MATLAB 9.4.

4.1. Parameter Specification

The number of neurons in a SLFN hidden layer defines the model complexity and has a significant impact on the underfitting/overfitting trade-off, unless some regularization is used. In ELM optimization, regularization parameters such as and controls the trade-off between model complexity and training error when solving the associated optimization problem. In GR-ELM and GOR-ELM, the parameter [0, 1] controls the sparsity of the network: choosing = 1 means that we prefer a sparse network over a dense network. These parameters are usually chosen using grid-search and a cross-validation method, such as k-fold, as done by Inaba et al. (2018), Zhang & Luo (2015), Deng et al. (2009), Huang et al. (2004) and others.

Note that some regularization parameters can have fixed values (e.g. 1) and the resulting minimization problem is equivalent, i.e., the result does not change if the problem is divided by a constant. According to Inaba et al. (2018), the value of in ADMM algorithms is usually fixed in = 1.

First, we followed some decisions of Zhang & Luo (2015): in OR-ELM, we used 20 iterations as its stopping criterion, fixed = 1 and chose C by using 5-fold and the grid .

N = 1000 for all training algorithms and datasets. When considering GRELM, we used primal and dual residuals as its stopping criterion, as suggested by Inaba et al. (2018), with = 10and = 10, we fixed = 1, = 1 and obtained and by using 5-fold cross-validation. We also chose R-ELM parameter by using the same method, and fixed = 0 and = 0.

When considering GOR-ELM, we fixed = 1, = 1 and obtained {0, 0.25, 0.5, 0.75, 1} and by using cross-validation. We used a combination of the stopping criterion proposed in Section 3.2 with the same values of and used by Inaba et al. (2018), and a large maximum number of iterations (= 1000), i.e., the algorithm stops when one of the criteria is met.

4.2. Regression with outliers

In this paper, we consider experiments with the objective of reducing the impact of initialization parameters (random weights), since our methods are non-deterministic. An experiment is defined as the following procedure: The weights between input and hidden layers (A) and the random biases () are randomly generated according to a uniform distribution in the interval [1]; The training subset is randomly permuted; A sigmoidal activation function in the hidden layer is used: h(x) = 1/(1 + exp()); SLFNs are trained using the algorithms and parameters obtained in 5-fold, and we obtain the average relative root mean square error (aRRMSE) metricand time for the training and testing subsets.

Considering =, with l = 1, 2, . . . , m, as the l-th of the m

targets of the problem, aRRMSE is defined as the mean of each target relative

root mean square error (RRMSE):

where ˆis the output of a SLFN with respect to , and ¯= (1/N) .

The experiments were run 100 times, using the parameters discussed in Section 4.1. The average (and standard deviation) values of aRRMSE obtained in these experiments are shown in Table 2, 3 and 4, for the uncontaminated datasets and with outlier ratios of 20% and 40%, respectively. The best results were highlighted in boldface. We also present a boxplot of the obtained results in Fig. 3, where points considered outliers were hidden for better visualization.

In all datasets, similar results were obtained with all training methods, when considering the uncontaminated dataset. In this case, as shown in Table 2, GOR-ELM achieved better metrics than the other algorithms in some datasets. Since the GOR-ELM main objective is to turn an SLFN robust to outliers, these results might imply that the datasets have some noise in its samples.

When considering that 20% of each training subset is contaminated with outliers, R-ELM and GR-ELM achieves worse results than the robust methods. This is expected, since they use the Frobenius (or ) norm, which are not suitable to deal with outliers. The better results, in this case, were obtained by robust techniques, with GOR-ELM winning in 12 of the 15 datasets. When considering 40% of outliers, similar results were obtained, with GOR-

ELM achieving better aRRMSE in 9 of 15 datasets.

It can be noted that in the aRRMSE values of GOR-ELM in the test sets remained similar in some datasets in the presence of outliers, when compared to the aRRMSE values obtained by the method when trained using the uncontaminated datasets. These results support GOR-ELM robustness to outliers characteristic.

In the presence of outliers, OR-ELM and GOR-ELM achieved better results, which was expected. Since its model considers relations between different targets (it uses matrix norms), GOR-ELM achieved better results than OR-ELM in most cases, showing that it can be a proper and robust technique in MTR tasks with the outliers inserted, according to the described methodology.

Table 5 show the time, in seconds, spent to train SLFNs with the respective algorithms. Our simulations show that GOR-ELM training stage was slightly slower than OR-ELM and GR-ELM in most cases, which was expected, since more optimization steps are needed in each iteration. Since all networks have similar architecture size, the time spent to test a set of data are almost the same.

Table 6 resumes the obtained parameters in the 5-fold cross validation for the tested methods. Table 7 shows the mean number of nodes after the training stage, and we can observe that the algorithms were not capable of pruning hidden nodes (or pruned a small number of nodes) with no impact on the model performance. This also suggest that better results may be obtained by increasing the number of neurons.

Fig. 3a shows the results of the tested methods when not considering outliers. In such cases, it is possible to see in the boxplots, the median of the values are almost the same for all methods (around 0.55). When considering outliers, the median becomes unbalanced, as showed in Fig. 3b and Fig. 3c, and robust techniques achieve smaller values than others, where GOR-ELM reaches the best one.

To confirm the statistical significance of our results, we use the Friedman test (Demˇsar, 2006). We considered a null hypothesis that the four compared

Table 2: Average aRRMSE with corresponding standard deviation for training and testing in MTR problems without outliers. Entries with indicates a value less than 10

Table 3: Average aRRMSE with corresponding standard deviation for training and testing in MTR problems with 20% of outliers. Entries with indicates a value less than 10

methods are equivalent against the alternative hypothesis that they are not, and a significance level of 10%. When considering uncontaminated datasets, the null hypothesis was not rejected, meaning that GOR-ELM is statistically equivalent to GR-ELM, R-ELM and OR-ELM.

However, when considering that the training dataset is contaminated with outliers, the null hypothesis is rejected with p-values of 0.000009 and 0.000271,

Table 4: Average aRRMSE with corresponding standard deviation for training and testing in MTR problems with 40% of outliers. Entries with indicates a value less than 10

Table 5: Average time (seconds) with corresponding standard deviation for training and testing in MTR problems. Smaller training times are highlighted in boldface. Entries with indicates a value less than 10

for outlier ratio of 20% and 40%, respectively. Thus, we can use a post-hoc test, such as the Nemenyi test (Demˇsar, 2006). The results of this test are shown in Fig. 4, where CD means the critical distance, i.e., the minimum spacing between two methods ranks so they can be considered statistically different, and methods

Figure 3: Boxplots of Table 2 (a), Table 3 (b) and Table 4 (c). The outliers were not shown for better visualization.

connected by a line are equivalent.

The results presented in Fig. 4 show that GOR-ELM is equivalent to ORELM and statistically better than GR-ELM and R-ELM, when the training dataset is contaminated with outliers.

4.3. Regression with outliers - Incremental approach

In this section, we compare IGOR-ELM only with IR-ELM in MTR tasks, since it showed better generalization performance than other incremental techniques, such as I-ELM and EM-ELM (Xu et al., 2016). We set for each method an initial number of hidden nodes ( ) equal to 100 for all datasets and 9 batches

Table 6: Parameter specifications for MTR datasets.

of 100 nodes were added (up to a maximum of 1000 hidden nodes). In this test, we also specified the parameters as explained in Section 4.1.

Similar experiments to those described in Section 4.2 were run 100 times, with the objective of reducing the random initialization impact of the methods, and the average (and standard deviation) values of aRRMSE obtained in these experiments are shown in Table 8, 9 and 10, for the original datasets and with outlier ratios of 20% and 40%, respectively. Similarly, the best results were highlighted in boldface.

The incremental results were similar to the ones presented in Section 4.2: When considering the uncontaminated dataset, IGOR-ELM and IR-ELM achieved

Table 7: Mean number of nodes of GR-ELM and GOR-ELM classifiers for MTR datasets.

close metrics, with the former reaching higher values in some cases. Considering the contaminated datasets, IGOR-ELM achieved even better results, when compared with IR-ELM.

It can be noted that in the tested incremental approaches, a large aRRMSE variation of both techniques occurred when comparing the achieved values in uncontaminated datasets with the obtained in contaminated ones. Since this did not occur in the non-incremental approaches, it can be due to the update step (Schur Complement) of the incremental methods.

In this experiment, IGOR-ELM was the only robust method, and thus, it obtained the best values of aRRMSE in the majority of datasets, when outliers were used in the training stage. Since it uses the GOR-ELM algorithm, the outlier structures were “detected” and the method showed robustness to them.

Figure 4: Critical distance considering aRRMSE in the testing subset and 20% (a) and 40% (b) of outliers in the training subset.

Table 11 show the time, in seconds, spent to train SLFNs with the respective algorithms. Our simulations show that IGOR-ELM training stage was slower than IR-ELM in most cases, which was expected, since it needs more optimization steps in each iteration.

As in Section 4.2, the results obtained for the incremental techniques have almost the same median (around 0.55) when trained without outliers, as we can see in Fig. 5. When considering contaminated datasets, the median of the results increase, where IGOR-ELM achieves smaller values than IR-ELM, which is not robust to outliers.

In our incremental approach, we use the Wilcoxon Signed Rank test (Wilcoxon, 1945) to confirm the statistical significance of our results. We chose this test since it is more suitable to compare two methods. We consider a null hypothesis that the distribution of difference between IGOR-ELM and IR-ELM metrics has median equals to zero, against the alternate hypothesis that it is less than zero, and a significance level of 10%.

When considering uncontaminated datasets, the null hypothesis was not rejected, meaning that both methods are statistically equivalent. However, when the training dataset has outliers, it was rejected with p-values of 0.0014 and 0.0017, for outlier ratios of 20% and 40%, respectively. This means that the distribution of IGOR-ELM metrics has a inferior median than IR-ELM. Thus, the former is statistically better than the latter, since inferior metrics in regression tasks implies in better performance.

Table 8: Average aRRMSE with corresponding standard deviation for training and testing incremental methods in MTR problems without outliers. Entries with indicates a value less than 10

5. Conclusions

In this paper, we proposed GOR-ELM and its incremental version (IGORELM), which extends the OR-ELM and GR-ELM algorithms to deal with multi-target regression problems. Instead of considering the Frobenius norm in the model error, we use the norm, which is more robust to outliers. For the proposed method, OR-ELM is a particular case of GOR-ELM, when the output has one dimension and when we set = 0 and = 1.

When we consider the non-incremental algorithms, our experiments showed that the proposed method achieved similar values of aRRMSE to those obtained

Table 9: Average aRRMSE with corresponding standard deviation for training and testing of incremental methods in MTR problems with 20% of outliers. Entries with indicates a value less than 10

by the other techniques, even when the dataset is not contaminated. However, its training stage was slightly slower. As expected, when we randomly inserted outliers in the training stage (20% and 40% of the samples), GOR-ELM usually showed better performance when compared to the other techniques. As the Friedman and Nemenyi statistical tests showed, GOR-ELM is better than the compared techniques in the presence of outliers.

When we consider the number of nodes, our experiments showed that GORELM was, in most cases, not capable of reducing the number of its nodes without compromising the model error. Thus, if the desired model error was not

Table 10: Average aRRMSE with corresponding standard deviation for training and testing of incremental methods in MTR problems with 40% of outliers. Entries with indicates a value less than 10

achieved, IGOR-ELM can be used to increase the node number of the network. We compared this technique with IR-ELM, which can add nodes to a SLFN trained using R-ELM. When considering the original datasets, similar results were obtained. However, when the datasets were contaminated, IGOR-ELM showed better performance than IR-ELM in most cases. The Wilcoxon signed rank test showed that IGOR-ELM was better than IR-ELM in tasks with outliers.

For future work, we can consider using other types of hidden nodes and activation functions, since we only tested the sigmoidal additive ones. We can

Table 11: Average time (seconds) with corresponding standard deviation for training and testing of incremental methods in MTR problems. Smaller training times are highlighted in boldface. Entries with indicates a value less than 10

also improve the network training time using GPUs or more efficient linear algebra libraries. Additionaly, we can also consider training GOR-ELM in a distributed approach, since ADMM can be trained using multiple processors, where each one handles part of a dataset.

Acknowledgments

The authors wish to acknowledge the support of the Brazilian Federal Agency for Support and Evaluation of Graduate Education (CAPES) through Grants

Figure 5: Boxplots of Table 8 (a), Table 9 (b) and Table 10 (c). The outliers were not shown for better visualization.

88881.062156/2014-01 and 1753509. F. Inaba was supported by a scholarship from the Brazilian Council for Development of Science and Technology (CNPq). P. Ciarelli thanks the partial funding of his research work provided by CNPq through grant 312032/2015-3. The work of E. O. T. Salles was support by Funda¸c˜ao de Amparo `a Pesquisa e Inova¸c˜ao do Esp´ırito Santo (FAPES) under Grant 244/2016.

Table 12: Parameter specifications of incremental methods for MTR datasets and its mean number of nodes.

References References

Akusok, A., Bj¨ork, K.-M., Miche, Y., & Lendasse, A. (2015). High-performance

Bartlett, P. L. (1998). The sample complexity of pattern classification with

Boyd, S. (2010). Distributed Optimization and Statistical Learning via the

Boyd, S., & Vandenberghe, L. (2009). Convex optimization volume 1.

Chen, C., He, B., Ye, Y., & Yuan, X. (2016). The direct extension of admm

Chen, C., Li, Y., Yan, C., Guo, J., & Liu, G. (2017). Least absolute deviation-

Demˇsar, J. (2006). Statistical comparisons of classifiers over multiple data sets.

Deng, W., Zheng, Q., & Chen, L. (2009). Regularized extreme learning machine.

Dˇzeroski, S., Demˇsar, D., & Grbovi´c, J. (2000). Predicting Chemical Parameters

Feng, G., Huang, G., Lin, Q., & Gay, R. (2009). Error minimized extreme

Ghosn, J., & Bengio, Y. (1997). Multi-Task Learning for Stock Selection. Ad-

Hodge, V. J., & Austin, J. (2004). A Survey of Outlier Detection Methodologies.

Huang, G., Huang, G.-B., Song, S., & You, K. (2015). Trends in extreme

Huang, G.-B., Chen, L., & Siew, C.-K. (2006a). Universal Approximation Us-

Huang, G.-B., Zhou, H., Ding, X., & Zhang, R. (2012). Extreme Learning

Huang, G.-B., Zhu, Q.-Y., & Siew, C.-K. (2004). Extreme learning machine: a

Huang, G.-B., Zhu, Q.-y., & Siew, C.-k. (2006b). Extreme learning machine:

Inaba, F. K., Salles, E. O. T., Perron, S., & Caporossi, G. (2018). DGR-

Kocev, D., Dˇzeroski, S., White, M. D., Newell, G. R., & Griffioen, P. (2009).

Liang, N., Huang, G., Saratchandran, P., & Sundararajan, N. (2006). A fast

Ma, Y., Li, C., Mei, X., Liu, C., & Ma, J. (2017). Robust sparse hyperspectral

Mandanas, F., & Kotropoulos, C. (2018). M-estimators for robust multidi-

Mart´ınez-Mart´ınez, J. M., Escandell-Montero, P., Soria-Olivas, E., Mart´ın-

Petersen, K. B., & Pedersen, M. S. (2007). The Matrix Cookbook. Citeseer,

Rao, C. R., & Mitra, S. K. (1971). Generalized Inverse of Matrices and Its

Spyromitros-Xioufis, E., Tsoumakas, G., Groves, W., & Vlahavas, I. (2016).

Tsoumakas, G., Katakis, I., & Vlahavas, I. (2010). Mining multi-label data.

Tukey, J. W. (1977). Exploratory Data Analysis. Addison-Wesley.

Wang, Y., Wipf, D., Ling, Q., Chen, W., & Wassell, I. (2015). Multi-Task

Wilcoxon, F. (1945). Individual Comparisons of Grouped Data by Ranking

Xu, W., Wang, M., Cai, J.-F., & Tang, A. (2013). Sparse Error Correction

Xu, Z., Yao, M., Wu, Z., & Dai, W. (2016). Incremental regularized extreme

Zhang, K., & Luo, M. (2015). Outlier-robust extreme learning machine

Zhang, Y., & Yeung, D.-Y. (2012). A Convex Formulation for Learning Task

Zhen, X., Wang, Z., Islam, A., Bhaduri, M., Chan, I., & Li, S. (2016). Multi-

Zhen, X., Yu, M., He, X., & Li, S. (2017). Multi-Target Regression via Robust

Designed for Accessibility and to further Open Science